So Ben Garrison or whatever gets a bunch of shares and comments for his bullshit and the robots pick up on it and try to recreate it, and then they pick up on the AI version, and it just keeps going until it's incomprehensible, all the while everyday Facebook users keep sharing it & it mutates
1 / 4
Comments
That sort of transformation requires an extraordinary energy source.
Otherwise, yeah, those are hallmark AI bots, and we are all the worse for it.
I have to ask...what is the actual motive? What is the profit model of getting a lot of likes on Facebook with automatic AI generated garbage?
Ai can get SO FUCKING DISTURBING.
uh, I forget the name of that prompt that just kept producing more and more horrific settings for basically the same woman's face, no matter what you did to it after a while.
sorry. I know it's Wrong, now. I shake my fist at AI and all its works!
But for a hot minute there...I lol'd.
And shuddered.
I replied earlier and referenced this article. And now I've found it!
1950s wife discovers lesbianism when her husband goes to war?
But maybe... to be expected? It can only work with what it already has.
This is the comic version of that.
If they want me to treat it as art I’m gonna critique it like art.
1. They can't count.
2. They don't understand syntax.
So the reason there are a *bunch* of coffee mugs is that that's what you get even if you just ask for *a* coffee mug.
https://hkotek.com/blog/text-to-image-models-shallow-parsing/
Excuse me?
which you get pretty much whether you're aiming for that or not.
Ironically they'd use significantly more power than the (minuscule) amount that they actually use if the creator actually cared about the quality, which he clearly does not, because he'd actually be doing selection & regeneration to fix errors.
Also, I just came across this 👇
https://futurism.com/the-byte/ai-datacenters-worse
old_man_yelling_at_clouds.png
This is realtime stable-fast pipeline + SDXL Turbo 512x512 batch 12 by Dan Wood on a RTX 4090. 77 frames per second. 1,6 milliwatt hours per image. On consumer gaming hardware, let alone server hardware.
Also, the source fo your source talks about data centre emisisons from 2020 to 2022. AI last year was only 10% of server usage. 2020 to 2022 it was a tiny fraction. Can't credit that to AI.
But carbon offset credits are an age-old debate that isn't at all datacentre-specific, let alone AI-specific.
Also [slightly moving my goal posts here], you know they're stealing intellectual property at a scale not known to mankind since ppl started to airbrush their hand prints at cave walls, don't you?
If a model is used one time or a trillion times, the training cost is the same.
.
—my dying words, probably